In [1]:
# Imports
from Models.model import DQNAgent
from IPython.display import Image
from Utility import wrappers
import matplotlib.pyplot as plt
import PIL
import imageio
import numpy as np
import pandas as pd
import gym
# Initializing the wrapped environment.
env = wrappers.wrap_deepmind(gym.make("BreakoutDeterministic-v4"))
In [2]:
# Print the action space and action's meaning.
print("Total number of actions:", env.action_space.n, "\n")
for i in range(len(env.unwrapped.get_action_meanings())):
print("Action", i, ":", "'{}'".format(env.unwrapped.get_action_meanings()[i]))
In [3]:
# Print the preprocessed starting state
state = np.array(env.reset())
state = np.expand_dims(np.array(state), axis=0)
print("Dimension of the preprocessed state:", state.shape)
In [4]:
"""Watch a random agent play breakout"""
# To test the agents's performance, we need a different wrapper for the environment,
# where the rewards are not scaled, and the episode resets on true game over.
env_test = wrappers.wrap_deepmind_test(gym.make("BreakoutDeterministic-v4"))
state = env_test.reset()
renders = []
score = 0
while(True):
renders.append(PIL.Image.fromarray(env_test.render(mode='rgb_array')))
_, reward, done, _ = env_test.step(env_test.action_space.sample())
score += reward
if done:
env.close()
break
file_path = "./gifs/random_agent.gif"
imageio.mimsave(file_path, renders, fps=30)
with open(file_path,'rb') as f:
display(Image(data=f.read(), format='png'))
print("Score achieved by random agent: ", score)
In [5]:
# Creating the agent
agent = DQNAgent(env)
In [ ]:
# 1st training session
# Training for 50k episodes
agent.run(num_episodes = 50000, mode = "train", weights_path = "./weights/model_weights_1", scores_path = "./logs/score_1")
In [6]:
# Training was interrupted
# Elapsed time: 45 hrs
# Epsison: 0.2 | Max. Score: 14 | Max. Average Score: 3.6
# plotting the scores
data = pd.read_csv("./logs/score_1").values.flatten()
agent.plot_scores(data)
In [7]:
# Loading the model's weights trained so far
agent.model.load_weights("./weights/model_weights_1")
# Resetting the epsilon
agent.reset_exploration(epsilon=0.25)
In [ ]:
# 2nd training session
agent.run(num_episodes = 30000, mode = "train", weights_path = "./weights/model_weights_2", scores_path = "./logs/score_2")
In [8]:
# Training was interrupted
# Elapsed time: 60hrs
# Epsison: 0.2 | Max. Score: 30 | Max. Average Score: 6.89
# plotting the scores
data = pd.read_csv("./logs/score_2").values.flatten()
agent.plot_scores(data)
In [9]:
# Loading the model's weights trained so far
agent.model.load_weights("./weights/model_weights_2")
# Resetting the epsilon
agent.reset_exploration(epsilon=0.05)
In [ ]:
# 3rd training session
agent.run(num_episodes = 30000, mode = "train", weights_path = "./weights/model_weights_3", scores_path = "./logs/score_3")
In [10]:
# Training was interrupted
# Elapsed time: 72hrs
# Epsison: 0.2 | Max. Score: 35 | Max. Average Score: 7.56
# plotting the scores
data = pd.read_csv("./logs/score_3").values.flatten()
agent.plot_scores(data)
In [11]:
# To test the agents's performance, we need a different wrapper for the environment,
# where the rewards are not scaled between -1 and +1
env_test = wrappers.wrap_deepmind_test(gym.make("BreakoutDeterministic-v4"))
# Create a new instance of the model (agent), for testing its performance.
agent_test = DQNAgent(env_test)
# Loading the best model weights trained so far.
agent_test.model.load_weights("./weights/model_weights_3")
In [12]:
"""Evaluation of the model"""
# Run the simulation
renders, score = agent_test.run(num_episodes=5, mode='test')
# Make a .gif file from the frames
file_path = "./gifs/intelligent_agent.gif"
imageio.mimsave(file_path, renders, fps=60)
# Watch an intellgent agent play breakout.
with open(file_path,'rb') as f:
display(Image(data=f.read(), format='png'))
print("Score achieved by Intelligent agent: ", score)
